Back to Guides
Enterprise Software 15 min read

The case for a working group instead of an AI governance committee

The case for a working group instead of an AI governance committee

One of the most-cited guides on standing up an AI governance committee contains an admission in the middle of it. OneTrust, describing its own structure, says the executive-level committee meets quarterly to review strategy, policy, and performance, and that the committee sets direction while smaller working groups handle day-to-day reviews. The committee is the layer above the work. The working groups are the work.

A company of 50 to 500 people that copies this model copies the quarterly layer and never staffs the part that decides anything. Avoiding that is not a matter of trimming a charter. It is a choice between two bodies with different lifespans, membership tests, and outputs.

What an AI governance committee is for

An AI governance committee is a standing, chartered body that sets AI policy, defines risk tolerance, reviews use cases above a risk threshold, keeps an inventory of systems in use, and reports upward. Its defining feature is permanence: it exists until dissolved, and its authority comes from a charter rather than from a specific decision.

The published seat lists are consistent. OneTrust names Legal, Ethics and Compliance, Privacy, Information Security and Architecture, R&D, and Product Management, with ten roles in its responsibilities table; the GSD Council’s version lists seven, adding an executive sponsor, a governance chair, and data governance. The recurring operating models — one central body approving everything, a federated split where business units self-approve low-risk use, or three lines of defense with internal audit on top — all presume standing functions to populate them.

That presumption is the problem, and it is separate from what governance looks like with no compliance function at all, which we take up in our look at governance without a compliance team. The question here is narrower: someone has to decide which tools are sanctioned and what may be pasted into them, so which body decides?

Where the committee template comes from

The template is not wrong. It is borrowed, and the lender is larger than you.

Read the most visible charters by their source rather than their advice. Two of them are university documents, covering teaching, research, clinical activity, and campus-wide tool inventories. Another is a municipal terms of reference, adopted by Prince Edward County in 2025, under which the committee advises the Chief Administrative Officer.

These are real documents, written by institutions with committee infrastructure, a general counsel, and a duty to deliberate in public. None was designed for a 200-person company where the person who would chair the committee also owns procurement and onboarding.

The arithmetic is unflattering. Seven named seats at a 200-person company puts three and a half percent of the company in one room, and at least four of those seats describe functions that company does not have. What arrives instead is one person wearing four hats and an invitation that gets moved twice and then cancelled.

What a committee is good at, and what it is bad at

A committee is good at three things, none of them trivial. It sets risk appetite, a judgment no individual should make alone. It resolves genuine conflicts between functions, where security wants a ban and sales wants a pilot. And it absorbs accountability for a refusal, so the person saying no to a senior colleague is backed by a body.

What it is bad at is deciding quickly and often. A standing review body creates a queue, a queue creates delay, and delay under deadline pressure produces a route around the queue — taken by the people with the most urgent work and the most standing to ignore a process.

That is not a soft cost. Bain’s decision-effectiveness research, a ten-year program covering more than a thousand companies, found that how well an organization makes and executes decisions correlates with its financial results at a minimum 95% confidence level, and just as strongly with employee attitudes.

Mid-market evidence suggests discussion is where most organizations already sit. In a Censuswide survey of 401 US IT leaders at organizations of 200 to 5,000 employees, published by Netrio in June 2026, 42% reported a formal AI policy with actively enforced controls and 53% full visibility into AI tool usage; 42% had a confirmed AI-related security incident or exposure in twelve months and 31% a near miss. Those are not numbers a quarterly meeting closes. They describe decisions lagging adoption.

Committee or working group: the honest comparison

A working group is a small, time-boxed body convened to produce named deliverables on a bounded problem, with a sponsor who holds the authority and a sunset date.

Dimension Standing committee Working group
Source of authority A charter A sponsor who can already sanction a tool and sign a policy
Membership test Function represented Has used the tools on real work in the last month
Size 7 to 10 seats 4 to 6 people
Cadence Monthly or quarterly, fixed Weekly for six weeks, then monthly, then on trigger
Primary output Minutes, policy, approvals A sanctioned-tool list, a data rule, a decision log, a training session
Lifespan Until dissolved Until its deliverables ship, with a renewal date
Right when Functions exist, risk appetite is contested, an auditor needs a chartered body Nothing is written down and the tools are already in use

The argument is not that governance matters less at this size. It is that a smaller company needs a set of decisions, not a body authorized to make them. The committee produces authorization; the working group produces decisions.

One failure belongs in neither column: conflating a board-level oversight body with a management-level approval body. Auditors expect the two kept apart, and at 200 people they collapse into the same four people anyway.

How to constitute a working group

Four to six people, one of whom is the sponsor, one of whom writes things down.

The sponsor is whoever can already act — usually the COO, VP of operations, or chief of staff. The test is whether they can sanction a tool, approve a purchase, and sign a one-page policy without asking anyone. If not, the group produces recommendations that wait on someone who was never in the room.

Every member has used the tools on real work in the last month. This is the rule no published seat list contains, and the one that decides whether the group is a control or theater. Those lists are drawn by function; none requires that a member has ever written a prompt. Someone who has not used the assistant cannot tell whether a proposed use case is a reasonable application or a bad idea with good slides, or when an output is wrong in the specific way these systems are wrong.

Include the person who would be blamed. If a client’s confidential material leaks, somebody’s name is on the engagement, and that person has better instincts about what must never be pasted into a chatbot than any job title does. Teaching the same judgment to everyone else is the subject of what a policy rollout has to be accompanied by.

Include one skeptic, not two. One person who thinks this is all risk improves the output. Two produce a ban, and bans move usage somewhere you cannot see, which is the case for sanctioning tools rather than prohibiting them.

Name a scribe in the first meeting — not a secretary, but the owner of the decision log. Nobody needs a seat whose contribution is a signature: a reviewer who confirms a document they have not read is a record, not a control.

The first six weeks

Six weekly sessions of sixty to ninety minutes, each ending in a written decision. The sequence matters because each week unblocks the next.

Week Question settled Artifact
1 What are we actually using? Inventory of AI tools and features in use, by owner, built by asking without penalty
2 Which tools are sanctioned, on which accounts? Sanctioned-tool list, naming the consumer logins to migrate
3 What data may go in? A three-tier data rule, specific to your data, on one page
4 Who reviews what before it leaves? Review rule by output type, naming the reviewer rather than the step
5 What happens when it goes wrong? Escalation path with a named first call and a penalty-free reporting route
6 How does anyone learn this? A dated training plan, and the renewal date for the group itself

By week six the group has produced the substance of an AI use policy without having drafted one as its first act — the usual order, and the reason most policies read as though written for a reader nobody has met. The full sequence from policy to training to evidence is laid out in our playbook on safe AI use.

Then it stops meeting weekly: monthly for twenty minutes on an agenda that already exists, plus a session triggered by an incident or a vendor change to data handling. The renewal date is the discipline. At six months the sponsor either renews the group against a new deliverable or stands it down and keeps the log — a legitimate outcome a committee cannot produce, because its charter has no end.

The decision log is the deliverable

One table, one row per decision, five columns: date, decision, who decided, why, review date.

It beats a policy document for three reasons. It stays current, because a row is cheaper to add than a policy is to amend. It handles exceptions, so a one-off approval becomes a dated row rather than a precedent nobody can find. And it answers what auditors and clients actually ask, which is not whether you have a policy but whether you have been applying one.

It is also the cheapest route to the evidence the frameworks describe. NIST’s AI Risk Management Framework, published as AI 100-1 in January 2023, states under GOVERN 2.1 that roles and responsibilities and lines of communication related to mapping, measuring, and managing AI risks are documented and are clear to individuals and teams throughout the organization. A log with named deciders satisfies that sentence better than a charter listing seats, because it shows the roles operating rather than existing.

A usable row is specific: “2026-11-03 — Sanctioned Claude and ChatGPT on company accounts for client-facing drafting, excluding anything under a named-client NDA — DJ, COO — administrative controls and data-handling commitments verified on both, consumer logins disallowed — review 2027-05-01.” Compare that with a policy line stating that employees shall use approved tools in accordance with company standards. Only one of them can be followed.

When you do need a committee

A chartered committee is the right instrument in specific conditions, and pretending otherwise would be selling you something.

Form one when a regulator, auditor, or major customer requires a named body with minuted decisions. When you pursue ISO/IEC 42001:2023 certification, which wants documented leadership roles. When risk appetite is genuinely contested between functions that both hold budget. And when your exposure runs through building or fine-tuning a model that reaches customers rather than through using commercial assistants on your own work — a distinction the EU AI Act draws by role rather than headcount, with deployer duties in Article 26 and a heavier regime for whoever places a system on the market under their own name.

Otherwise the working group is not a lightweight committee. It is the correct body, and the committee is what you add if that list starts applying to you.

One obligation holds either way. The EU AI Act’s Article 4 AI-literacy provision has applied since 2 February 2025, and NIST’s framework treats personnel training as a control in its own right. Neither cares what your governing body is called; both expect the people operating these systems to be competent, which is the thread running through the argument that buying licenses is not the same as building capability.

What it costs to run

In hours: roughly 30 to 50 person-hours across the six sessions, then one to two hours a month, with the sponsor carrying about a third. In money: a few hours of outside counsel to confirm which rules reach you, the delta between consumer and business accounts for the tools you sanction, and the cost of building the fluency the membership rule assumes. Applied AI enablement sessions run in a market range of roughly $2,000 to $15,000 depending on scope and headcount.

Frequently asked questions

What is an AI governance committee?

A standing, chartered body that sets AI policy, defines risk tolerance, reviews higher-risk use cases, maintains an inventory of AI systems, and reports to leadership or the board. Published seat lists run from seven to ten roles across legal, compliance, privacy, security, data, and the business.

Do we need an AI governance committee at 200 employees?

Almost certainly not. You need the decisions a committee would authorize, which a four-to-six-person working group with a sponsor who can act produces in about six weeks. Form a committee when an auditor, regulator, or customer requires a chartered body with minutes.

What is the difference between an AI governance committee and a working group?

A committee is permanent, authorized by a charter, and staffed by function. A working group is time-boxed, authorized by a sponsor who already holds the authority, staffed by people who use the tools, and judged on named deliverables with a sunset date.

Who should sit on an AI working group?

Four to six people: the sponsor who can sanction a tool and sign a policy, the person accountable if client material leaked, one informed skeptic, a scribe who owns the decision log, and one or two people doing the daily work. Every member must have used the tools recently.

Does a working group satisfy an auditor or a client security questionnaire?

A dated decision log with named deciders usually satisfies them better than a charter does, because it evidences controls operating rather than existing. GOVERN 2.1 asks for documented roles and clear lines of communication, not for a named body.

How do we stop the working group from becoming a committee?

Give it a renewal date. At six months the sponsor either renews it against a new deliverable or stands it down and keeps the log. Permanence without a deliverable is how a queue forms.

Should our AI policy come before or after the working group?

After. A policy drafted as the first act is written by whoever volunteers, for a reader nobody has met. Run the six weeks, then write the policy from decisions already being followed.

What if nobody on the group has used the tools?

Fix that before the first session. A body that cannot evaluate an output cannot evaluate a use case, and a review step staffed by people who do not know where these systems fail is a signature, not a control.

Key takeaways

  • The committee template is borrowed from institutions with standing functions. Even one of its best-known advocates describes a committee that meets quarterly while working groups do the reviews.
  • A committee is good at setting risk appetite, resolving conflicts between functions, and absorbing accountability for a refusal. It is bad at deciding quickly and often, which is the job you have.
  • Mid-market numbers show decisions lagging adoption: 42% with actively enforced controls, 53% with full visibility into tool usage, and 42% with a confirmed AI-related incident in twelve months.
  • Staff four to six people under a sponsor who can already act, and apply the rule no published seat list contains: everyone has used the tools on real work.
  • Six weekly sessions produce an inventory, a sanctioned-tool list, a data rule, a review rule, an escalation path, and a training plan. The policy is written from those decisions, not before them.
  • The decision log is the deliverable. Dated rows with named deciders evidence governance operating, which is what GOVERN 2.1 and a client questionnaire both ask for.

If you want a second opinion on which body your company needs, book a free AI-readiness call. We will look at what your team already uses, say whether a working group or a chartered committee fits, and name the decisions waiting on a meeting.

Last Updated: Nov 16, 2026

DJ

Dirk Jan van Veen, PhD

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles